Skip to main content

How this Helps

Visual Search finds near-duplicates, repeated patterns, and conceptually similar content in your dataset at the image, object, or region level. Use it for anomaly investigation, pattern clustering, and accelerated QA review.

​
Overview

Visual Layer runs similarity search across your dataset using deep visual embeddings. You can compare against any of the following references:
  • An entire cluster representative
  • An image preview, from any place an image appears in the UI
  • A cropped region drawn directly on an image preview
  • An object bounding box inside an image
  • An external image uploaded from your device
Every image preview in the product supports two similarity actions on hover:
  • Find Similar — searches against the whole image.
  • Crop to find similar — opens an inline cropper on the thumbnail so you can draw a Region of Interest and search against that specific area. This is the recommended path when you want to isolate a defect, object, or detail rather than match the full frame.
Because both actions live on the preview itself, you never need to leave the current view to start a visual search.

​
Similarity Search on a Cluster

Use this path to find clusters visually similar to one you already see.
  1. Hover over the target cluster card in the Dataset Exploration view.
  2. Click Find Similar at the bottom of the card.
The search runs against the cluster’s representative embedding and returns the most similar clusters in the dataset.

​
Similarity Search from an Image Preview

Any image thumbnail in the product surfaces the same two similarity actions. The entry points differ only in where you are when you reach for them.

​
Entry points

Hover an image thumbnail in any of these locations:
  • On a cluster card in the exploration view. Cluster cards show a preview carousel of representative thumbnails. Hovering a single thumbnail reveals the similarity actions, so you can launch a search against that one image without opening the cluster first.
  • Inside an opened cluster. Click into a cluster to view its full grid, then hover any thumbnail in the grid.
  • On the image details page. Open an image to its details view and hover the main image.

​
Search against the whole image

  1. Hover the image thumbnail.
  2. Click Find Similar.
The search returns images whose embeddings are closest to the full reference image.
Similarity search triggered from a hovered image thumbnail

​
Search against a Region of Interest

Region-of-interest search compares a cropped area of the reference against the dataset. Use it to isolate a specific defect, object, or pattern while ignoring background content.
  1. Hover the image thumbnail.
  2. Click Crop to find similar. An inline cropper opens directly on the thumbnail.
  3. Drag to draw the region you want to search against.
  4. Click Find Similar inside the cropper overlay to run the search.
The crop flow is available from every entry point listed above — the cluster card preview thumbnail, the opened cluster grid, and the image details page — so you can pick the access point that keeps you closest to your current context.

​
Similarity Search from an Object

When an image contains detected objects, each bounding box is its own similarity anchor.
  1. Open a cluster or image that has object detections.
  2. Hover directly over the object’s bounding box.
  3. Click the Find Similar icon that appears on the box.
The search returns images that contain visually similar objects rather than similar full scenes.

​
Similarity Search by External Image Upload

Use this path when your reference image is not already in the dataset.
  1. On the Dataset page, click the Visual Search icon near the main search bar.
  2. Upload an image from your local device.
  3. (Optional) Drag to crop the uploaded image before searching.
  4. Click Find Similar to run the search against your dataset.
Similarity search via external image upload

How Search and Filter Work

The embedding theory behind visual similarity and nearest-neighbor search.

How to Search & Filter

Step-by-step guide to running searches and filters, plus the full list of available options.

Exploration Recipes

Step-by-step workflows that combine visual search with filtering.

Saved Views

Persist a search so you and your team can return to it later.